Power-system dynamic simulation gets a fast stand-in: fine-tuned Chronos-2 beats a traditional AI model on error and runs 563 times faster
Synopsis
The pretrained time-series foundation model Chronos-2 was tested as a surrogate for power-system dynamic simulation. On 1,800 load-disturbance samples from an IEEE 9-bus system, retrieval-augmented in-context learning beat random reference selection, and fine-tuned Chronos-2 had lower errors than a traditional spatiotemporal graph neural network baseline on the 178 held-out samples, with power-balance regularization improving every reported metric except voltage-angle MAE. Its 35.5 ms inference was 563 times faster than the 20 s ANDES simulation.
Interpretation
A pretrained time-series foundation model can take on post-disturbance power-system trajectory simulation without training a dedicated model from scratch for each system. Earlier machine-learning surrogates were typically tailored to specific systems and operating conditions, and whether a general time-series foundation model could reproduce the physically constrained dynamics of a multi-bus system had not been verified. On 1,800 samples spanning six load-disturbance modes of the IEEE 9-bus system, fine-tuned Chronos-2 showed lower errors than the conditional spatiotemporal graph neural network baseline on voltage magnitude, voltage angle, frequency and rate of change of frequency across the 178 held-out samples.
Writing grid physics into the fine-tuning objective improves both point and probabilistic predictions of the foundation model. Standard fine-tuning keeps the native multi-quantile loss, while physics fine-tuning adds a penalty on active- and reactive-power balance residuals computed through the AC power-flow equations. The physics fine-tuned variant was superior to standard fine-tuning on every reported metric except voltage-angle MAE, with voltage-angle CRPS of 0.198, frequency CRPS of 0.037 and ROCOF CRPS of 0.028 versus 0.235, 0.039 and 0.030 for standard fine-tuning.
The pretrained weights themselves are a useful starting point rather than an incidental initialization. To isolate the effect of pretraining, the same Chronos-2 architecture was retrained with the same recipe but random initial weights. The randomly initialized control failed to converge on all attempts, while fine-tuned Chronos-2 outperformed the traditional AI model trained from scratch on the held-out set.
Perspective
The demonstrated scope is load-type disturbances on the IEEE 9-bus system (three generators, nine buses, six load points), with data generated by ANDES, the open-source Chronos-2 model and 178 held-out test samples; larger systems, cross-topology and cross-disturbance generalization, and power-system-specialized foundation models are listed as future work.
The gain of physics fine-tuning over standard fine-tuning does not hold for voltage-angle MAE, where the difference is small; the randomly initialized control failed to converge on all attempts, and the convergence criterion is not detailed; all conclusions come from a single system and a single model.